Figure 2 provides a series of modified scatter plots for the 10-year average homicide rates across the four sources. Correlations between WHO rates and those from UNODC and WB WDI ranged from 0.79 (1990-1999, 2000-2009) to 0.85 (2010-2018). The WHO average homicide rate was significantly different from the UNODC and WB WDI rates for the 2000-2009 and 2010-2018 periods. WHO and GHO homicide rates were correlated at 0.77 (2001-2019) and 0.84 (2010-2018). The differences were statistically significant in both 10-year periods. Therefore, while the homicide rates from each source tend to trend together, the differences between their overall averaged rates are significant.

The UNODC and WB WDI rates were perfectly correlated across each of the 10-year periods. Like the 5-year averages, a few nations have 10-year average homicide rates that differ slightly between the two sources. The average homicide rates for UNODC nations were not significantly different from their counterparts in the WB WDI in any of the 10-year time frames.

The GHO and the UNODC and WB WDI homicide rates were nearly perfectly correlated, at 0.98 (2001-2009) and 0.99 (2010-2018). As with the 5-year averages, however, in each of the 10-year periods the average GHO rates were significantly different in magnitude from the UNODC and WB WDI rates.

Figure 3 provides the best fit line representation of the models, along with the slope coefficient and p-value. These results are shown in tabular format in Table A3 of our Online Appendix. Figure 8 provides evidence that scholars would come to different conclusions for the effects of infant mortality, sex ratio, unemployment, and the education index on homicide rates depending upon the homicide data source they selected. During the 1990-1999 period, assuming the use of a two-tailed test, poverty was significantly associated with homicide rates when using WHO data but not with UNODC and WB WDI data (though it was significant when using the latter two data sources if employing a one-tailed test, which is not uncommon in this literature given theoretical expectations for many variables).

During the 2000-2009 period, sex ratio was not associated with homicide rates when using WHO and GHO data, but it was associated with homicide for UNODC and WB WDI data if a one-tailed test was used. For unemployment there were differences by data source in the 1990-1999 and 2000-2009 periods. In each period unemployment was not significantly associated with homicide rates when using WHO data, but it was significantly associated with homicide when using UNODC, WB WDI, and GHO data (though in the 2000-2009 range the associations were significant only when using a one-tailed test). There were multiple differences by data source for the education index. In all three 10-year periods, the education index was not significantly associated with homicide rates when using WHO data. In 1990-1999 and 2010-2018, however, it was significantly associated with homicide when using UNODC and WB WDI data (for the 1990-1999 period it was significant only with a one-tailed test). It was also significantly associated with homicide using GHO data in 2000-2009 (one-tailed test) and 2010-2018.

Figure 4 displays results when testing if the magnitude of the slope varies significantly by homicide data source. That is, is the difference in the slope coefficients for each variable significantly different from zero across all possible comparisons of data sources? In the 1990-1999 plot, sex ratio and unemployment both had differences in the slope coefficient when using WHO data relative to UNODC and WB WDI data. For sex ratio, however, once we applied a Bonferroni correction for multiple comparisons these differences were no longer significant.